The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The create_worker_agent function in SKILL.md accepts a task dictionary and interpolates several fields, including task_instructions, acceptance, and name, directly into a generated SKILL.md file for worker agents. There is no evidence of sanitization, escaping, or boundary markers to prevent the task_instructions from containing malicious prompt injection that could override the worker agent's intended behavior.
Ingestion points: The task object passed to create_worker_agent (specifically the task_instructions, acceptance, and name fields).
Boundary markers: None identified; the values are substituted directly into the template via string replacement.
Capability inventory: The skill spawns autonomous subagents using the Task tool, writes files to the local file system (write_file), creates directories (os.makedirs), and removes directories (shutil.rmtree).
Sanitization: None identified; placeholders like {task_instructions} are replaced with raw string values.
[DYNAMIC_EXECUTION]: The skill implements a runtime agent factory that generates executable agent skill files (SKILL.md) by substituting task-specific data into markdown templates. These generated skills are subsequently executed by spawning subagents through the Task tool in the dispatch_workers function. While this is the primary purpose of the skill, the incorporation of unsanitized external input into the generated executable content poses a security risk.

agent-factory